particlefuture/MCPDiscovery

> **MCP of MCPs** — automatically discover and configure MCP servers on your machine (remote or local).
After setup, you can usually just say:
> “I want to perform . Call the deep_search tool and follow the outlined steps.”
The goal is that you only install **this** MCP server, and it handles the rest (searching servers, selecting servers, configuring servers, etc.).
Demo video 🎥: https://youtu.be/W4EAmaTTb2A
Quick Setup
Choose **one** of the following:
1. **Remote** (simplest & fastest ⚡💨)
2. **Local (prebuilt)** — **Docker**, **uvx**, or **npx**
3. **Local (from source)** — run this repo directly
1) Remote 🌍⚡💨
Use the hosted endpoint (recommended for the simplest setup).
**Docs + guided setup:** https://mcp.1mcpserver.com/
Configure your MCP client
Add the following entry to your client config file:
* **Cursor**: ./.cursor/mcp.json
* **Gemini CLI**: ./gemini/settings.json (see Gemini docs)
* **Claude Desktop**:
* macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
* Windows: %APPDATA%\Claude\claude_desktop_config.json
* **Codex**:
* macOS: ~/.codex/config.toml
* Windows: %USERPROFILE%\.codex\config.toml
**Remote config (JSON):**
If you already have other servers configured, just merge this entry under mcpServers For example:
**Tip:** If your client supports it, move the config file into your **home directory** to apply globally.
---
2) Local (prebuilt) 💻
Use this when you want everything local, or when your MCP client only supports **STDIO**.
2A) Docker 🐳
Running on other host ports:
Running with stdio instead of streamable-http (You might see some delays when trying to connect):
2B) npx 📦
---
3) Local (from source) 🧩
Clone this repo and run directly.
> If your client supports remote url servers, you can use the **Remote** setup instead.
Optional: grant file-system access 📁
If you want your LLM to have file-system access, add an MCP filesystem server and point it at the directory you want to allow:
---
Architecture 🧠
There are two search modes:
Quick Search ⚡
For explicit requests like: “I want an MCP server that handles payments.”
Returns a shortlist of relevant MCP servers.
Deep Search 🌊
For higher-level or complex goals like: “Build a website that analyzes other websites.”
The LLM breaks the goal into components/steps, finds MCP servers for each part, and if something is missing, it asks whether to:
* ignore that part,
* break it down further, or
* implement it ourselves.
Deep Search stages:
1. **Planning** — identify servers, keys, and config changes
2. **Testing** — verify servers (via test_server_template_code)
3. **Acting** — execute the workflow using the configured servers
---
Change Log 🕒
* July 31 2025: Upgrade to 0.2.0. Added agentic planning.
* Dec 12 2025: Support for Gemini + Codex